commas inside numbers and lists
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 12 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron reacts to commas that split up big numbers like 65,000 or place names like Newark, DE. When it is turned down, it is most active on commas that separate items in a list of actions. Turning it up did not produce a clear change in topic.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
. This one, called iBART, was a thing of beauty. Free, too. It was written by two former high-school buddies—Ian Leighton, a sophomore at UC Berkeley, and David Hodge, a sophomore at the University
-
pelier, Vermont, on December 26, 1837. A hero of the Civil War and Spanish-American War, he was the first, and to date only, person promoted to Admiral of the Navy.⏎Marihuana
-
five-a-day, it would now seem that even that is not enough.⏎Nutritional boffins at the University College London surveyed 65,000 people over 12 years and concluded that we should actually be
-
nervous system, causing many of the disease’s symptoms.⏎How common is Citrullinemia Type 1?⏎Scientists estimate that 1 in 57,000 births are affected by Citrullinemia type I.⏎How
-
really learn.⏎Beers, Kylene and Robert E. Probst. Notice & Note: Strategies for Close Reading. Portsmouth, NH: Henemann, 2013.⏎Using the work of Louise
-
three Stanford researchers that delves into those pitfalls appears in Proceedings of the National Academy of Sciences. For their work, the team analyzed the results of 50,000 studies in 22 fields of science. “I think that
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: I have been on the internet, it's been a lot of social media. I've been working in the dark hours of work, I have been on a lot of social media, I have been on a lot of work. I have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
this product is for A SINGLE USER ONLY. That means that copying or sharing any part of this product with another teacher, grade level, school, department, district,
-
, engines, or generators to control and adjust production of biofuels or biomass-fueled power.⏎- Operate equipment to start, stop, or regulate biomass-fueled generators, generator units, boilers, engines,
-
||fortran|⏎This entry (in part or in whole) was last reviewed on 4 February 2012.⏎Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version
-
ctman & Horowitz, 2003). Studies find that people do not treat computers as people outside the mediation context (Goldstein, Alsio, & Werdenhoff, 2002) – just as people do
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L10:667 suppresses commas inside lists of names
- L10:74 suppresses
- L6:620 suppresses "and" in lists of three or more
- L8:1605 excites
- L9:77 excites commas in citations and name lists
- L9:738 suppresses
- L7:866 suppresses
- L7:288 excites
- L11:154 suppresses commas after citation numbers
- L8:2014 excites
- L8:1046 weakly excites
- L8:1652 weakly suppresses
- L10:1953 weakly suppresses
- L11:807 weakly suppresses
- L3:661 weakly excites
- L10:2676 weakly suppresses
- L4:3 weakly excites commas after citation fields
- L9:633 weakly excites
- L5:16 weakly excites
- L11:876 weakly excites 'other' after 'among' in lists
- L0:2986 weakly suppresses
- L0:2436 weakly suppresses
- L9:509 weakly suppresses
- L10:1643 weakly excites
Feeds later layers
- L13:977 strongly suppresses
- L13:758 suppresses 'and' joining a pair of options
- L16:1388 suppresses
- L14:1054 suppresses
- L15:677 suppresses
- L13:517 suppresses comma closing a list item
- L14:874 suppresses
- L15:611 suppresses
- L13:125 suppresses
- L14:1804 suppresses "and" in subject-area lists
- L15:545 suppresses
- L21:735 suppresses 'or' in ranges and alternatives
- L17:572 suppresses
- L16:1294 suppresses
- L14:2054 suppresses
- L14:1215 suppresses
- L16:2733 suppresses
- L16:1234 suppresses
- L14:2961 suppresses
- L13:900 weakly suppresses
- L19:1369 weakly suppresses
- L16:774 weakly suppresses
- L14:1746 weakly suppresses
- L15:631 weakly suppresses
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- inction
- .Dict
- owing
- either
- แดน
- 既
- 大片
- .ByteString
- umbing
- 初
Suppresses
- reas
- uku
- rine
- quito
- 和他的
- aches
- ことです
- chas
- aru
- lement